SPB Git

spb/wp31_uqo

Public
921 B · 57 lines yaml
1name: p4_diag_b8only_w642seed: 13mode: prisme4model:5  name: tiny_p46  vocab_size: 1287  d_model: 1928  n_layers: 49  n_heads: 610  n_kv_heads: 211  ffn_width: 51212  max_seq_len: 32013  w: 6414  ffn_w: 19215  n_slots: 1616  n_editor_heads: 217  r_max: 818init:19  backbone_checkpoint: experiments/runs/p4_tiny_backbone_seed01_20261029232626/checkpoint20  freeze_backbone: true21  auto_readout_scale: true22data:23  kind: synthetic24  tokenizer: char25  splits_dir: data/splits/synthetic_v126  loss_span: answer27  eval_splits:28  - dev_iid29  - dev_template30  - dev_depth31  - dev_compose32budgets:33- 834budget_probs:35- 1.036budget_per: position37train:38  batch_size: 3239  seq_len: 32040  steps: 300041  lr: 0.00142  warmup: 10043  weight_decay: 0.144  grad_clip: 1.045  log_every: 5046  eval_every: 50047  eval_batches: 1248  ckpt_every: 150049em_budgets:50- 051- 252- 853em_max_items: 40054model_overrides:55  readout_zero_init: false56  readout_init_std: 0.257
Files · main
Menu

Theme